RecommendationPaige describes using Google AI Edge Gallery with Gemma 4 for on-device tasks: describing images, transcribing audio in multiple languages, testing function calling, and using skills like building games, writing haikus, weather queries, and calendar scheduling.
AI Engineer · Research to Reality with Google DeepMind — Paige Bailey, Google DeepMind
RecommendationPaige recommends loading Gemma 4 directly in the browser via WebAssembly (sandboxed) and running it locally with Transformers.js.
AI Engineer · Research to Reality with Google DeepMind — Paige Bailey, Google DeepMind
RecommendationPaige recommends using Google AI Edge Gallery to run Gemma 4 on Android and iOS, analyze on-device model capabilities, and take advantage of free skills with locally installed models.
AI Engineer · Research to Reality with Google DeepMind — Paige Bailey, Google DeepMind
Recommendation · MoneyUse a coding agent like Claude Code to pull down your failing traces, modify your prompts, and improve your application instead of hand-writing fixes. Every change it proposes should trace back to a specific failure explanation — nothing should be guessed.
AI Engineer · From Vibes to Production: Evaluating and Shipping AI Agents That Work 101 — Laurie Voss, Arize AI
Recommendation · MoneyRun evals directly against the audio using audio-model-based evals, not just LLM-as-a-judge on transcripts. Flag interruptions, transcription accuracy/drift, and task success as voice-specific eval dimensions. You can create audio evals conversationally — e.g., ask Claude to create an audio eval for sentiment analysis using GPT audio that catches frustrated user tone on audio spans. Under the hood, the eval CLI queries spans, does trace analysis, filters, creates the eval, deploys it, and runs it against traces and spans in your actual system.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Recommendation · MoneyIntroduce a truncation algorithm for tool results — the speaker shipped one, stood it up in dev, tested against failed traces, and showed how latency decreased.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Recommendation · MoneyCreate an audio eval through Arize skills by prompting Claude — e.g., 'create an audio eval for sentiment analysis using GPT audio and catch frustrated user tone on those audio spans' — and the CLI queries spans, does trace analysis, then creates, deploys, and runs the eval against traces and spans.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Recommendation · MoneyFlag voice agent sessions that break time-to-first-audio SLAs; these SLAs remain important for customer service use cases.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Recommendation · MoneyBeyond latency, flag interruptions, accuracy issues, transcription drift, and task success in voice agent sessions.
AI Engineer · The Transcript Looked Fine. The Call Wasn't. — Debugging Voice Agents, Arize
Method · MoneyTo add price filtering to the app, ask the agent to add price filtering ('Let's add price filtering to this app.'); even if the agent doesn't find the price filtering problem in existing traces, the plan is to get it to add the price filtering feature; the speaker then asked the app 'Show me all toys under $9' to verify the price filtering fix, and it returned all toys under $9.00 with input and output visible in the trace response.
Method, not a verified protocol.
AI Engineer · From Vibes to Production: Evaluating and Shipping AI Agents That Work 201 — Laurie Voss, Arize AI
Recommendation · MoneyTrain SFT models only on trajectories that the verifier marked as passed (successful trajectories), filtering training trajectories by verifier pass/fail before SFT to raise model quality.
AI Engineer · Your LLM Judge Is a Confident Liar: Building Better Verifiers — Browserbase
Recommendation · MoneyThink of queries per second at the engine level — how many queries per second you can handle per replica of your inference engine — rather than at the user level, because aggregate demand comes from users.
AI Engineer · What Is an Inference Engine, Anyway? — Charles Frye, Modal
Recommendation · MoneyCharles Frye warns that one of the most common model quality bugs occurs at model release: tokenizers are often slightly bugged — they both tokenize and apply chat templates, and there are often issues there. They get fixed within a couple weeks after model release, but you always have to be on guard for it. His hot tip: set your engine up so that token IDs are logged.
AI Engineer · What Is an Inference Engine, Anyway? — Charles Frye, Modal
Recommendation · MoneyCharles Frye recommends starting to learn inference engines with MiniSGLang and NanoVLM — simple implementations of the engine architecture by the two teams, great for humans and incredible for agents because they can fit all the context in their context window. He also recommends reading Alexa Gordich's walkthrough blog of VLLM and working through it in a notebook, putting your coding agent on the codebase and asking questions, and using Cognition's DeepWiki, which crawls repos (including SGLang and VLLM), writes architecture diagrams and wiki-style pages, and lets you ask their agent questions — a nice pairing with your own agent setup.
AI Engineer · What Is an Inference Engine, Anyway? — Charles Frye, Modal
Recommendation · MoneyBlaine Billingsley describes pitching design work as a simple story about a person with a problem that the product will make better, rather than leading with metrics or business impact; he notes this is naturally how he thinks about product and software design, and that it kept him on the Gmail team a long time.
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet
Recommendation · MoneyBlaine recommends that when a team is stuck disagreeing between two ideas, they should go get 10 more ideas and find something in there — it relaxes the stress of the standoff.
medium
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet
Recommendation · MoneyWhen someone is pushy about their idea, Blaine Billingsley recommends saying 'cool, let's go for it, let's see what it looks like' and actually trying it; he also discourages binary 'should we do this or that' conversations because they turn into zero-sum games that make anyone pushy.
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet
Recommendation · MoneyIn meetings where nobody understands the project line items, Blaine Billingsley recommends saying 'I have no idea what this means' to break the ice; he acknowledges admitting this is vulnerable but calls it the way to break through unclear communication, and among the most powerful things you can do in those moments.
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet
Recommendation · MoneyWhen you hear an undefined jargon phrase, Blaine Billingsley recommends pushing for a concrete definition: ask what it means, offer interpretations, and make it mean something specific — an antithetical approach he finds really useful.
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet
Recommendation · MoneyWhen defending a risky decision out loud — like a launch delay concern — Blaine Billingsley recommends naming the tension and stopping; more often than not the tension drops and real problem-solving begins. He attributes the technique of 'labeling' — naming the other party's emotion or concern — to Chris Voss, former FBI hostage negotiator, noting it works as well on stakeholders as across a negotiating table.
Design Better · Blaine Billingsley: OpenAI designer on the collaboration problem AI hasn't solved yet